Moving From ‘Hallucinations’ to ‘Fabrications’

As the field of artificial intelligence continues to evolve and expand into various domains, including health care, a pressing issue has emerged that requires immediate attention from the psychiatric community. The term “hallucination” has been increasingly used to describe AI-generated content that is factually incorrect or exaggerated, particularly in health care contexts (Maleki, et al., 2024).

However, this terminology is problematic, as it conflicts with the specific scientific meaning of hallucinations in the mental health context. Hallucinations are sensory experiences experienced by an individual, and as such are false perceptions that can be at times bizarre, distressful, and disturbing to the individual. The underlying conditions that typically lead to an individual experiencing hallucinations could be psychiatric, neurological, or substance-induced. Using the word hallucinations in the context of AI diminishes the importance of this symptom and the pain it creates for those suffering from mental illness.

Unfortunately, the literature abounds with articles regarding AI hallucinations, and it is difficult to sway the technology community to abandon this term (Maleki, et al., 2024; Ji, et al., 2023). The field of psychiatry needs to champion alternative terms that are less likely to conflict with established mental health terminology. We support a unified definition of “AI hallucinations” as “Responses confidently providing fluent but factually incorrect or exaggerated responses, stemming mainly from statistical text prediction from training the AI models using potentially biased data and the stochasticity inherent in some AI algorithms rather than genuine understanding or reasoning.” We also suggest replacing the term “hallucination” with “fabrication” to differentiate this meaning from a hallucination in the mental health context.

AI fabrications pose significant risks, including the dissemination of potentially false health-related information (Linfield & Parsonnet, 2023). This is particularly concerning in health care contexts, where AI tools are increasingly being used to provide medical information and advice to patients. The medical informatics community must develop a robust architecture to prevent the spread of such misinformation and ensure the accuracy and trustworthiness of AI-generated content. Robin Emsley has eloquently written about the danger of a future filled with fictitious information (Emsley, 2023).

Beyond changing terminology, we propose a framework for addressing this issue, which includes evaluating AI-generated content against ground truth, implementing verification mechanisms that are human-reviewed, and ensuring citations and references to every claim. We also suggest prioritizing correctness, consistency with inputs, and trustworthiness in health care contexts, while limiting creativity and the use of anthropomorphic references.

This is not a trivial task, and there are no easy solutions. However, it is essential that the psychiatric and medical informatics communities take an active role in developing solutions to address this issue. The risks associated with the dissemination of false, nonfactual health-related information by AI tools are too great to ignore, and it is our responsibility as health care professionals to ensure that AI-generated content is accurate, reliable, and trustworthy.

We must work collaboratively to develop a methodology to prevent the spread of misinformation and ensure the accuracy and trustworthiness of AI-generated content in health care contexts. Psychiatry has a longstanding track record of developing taxonomies and definitions, and we should take the lead in this area. The medical informatics community likewise must develop a robust architecture to prevent the spread of misinformation and ensure the accuracy and trustworthiness of AI-generated content.

We urge psychiatrists to take an active role in addressing this issue and to work collaboratively to develop solutions that prioritize correctness, consistency with inputs, and trustworthiness in health care contexts. ■